Tencent Hy4 Preview: Open 770B Model Targets Coding and Research

What Tencent released

On Friday, August 28, 2026, Tencent posted a preview of Hy4, a new open-source AI model, on Hugging Face. The company is pitching it at software engineering, research and financial analysis.18 Hy4 preview uses a mixture-of-experts design with 770 billion total parameters. Only about 49 billion of them are used for any single text request.11 The context window is about one million tokens.17 Tencent calls it the most capable model it has built and says it is the biggest jump it has made between two generations of models.17

The weights were published on Hugging Face, ModelScope, GitCode and CNB. The model can also be reached through OpenRouter.5 Developers can use Hy4 preview through Tencent Cloud, OpenRouter and Tencent's own products.17 Several outlets report the license as Apache 2.0, which allows anyone to download, modify, fine-tune and use the model commercially.4915 One early write-up was less specific and simply said the license terms were on the Hugging Face model card.5 The bulk of the coverage, along with the summary of the model family's licensing, points to the permissive Apache license.2

The Reuters wire story, which was syndicated widely, focused on the scale and the target uses. It also noted that Tencent plans to deploy the model alongside its CodeBuddy and WorkBuddy products.1320

Why this release stands out

Much of what matters here is the jump in size. Hy3, the full model Tencent released on July 6, 2026, has 295 billion total parameters, 21 billion of them active, and a 256,000-token context window.3 Hy4 preview has more than twice as many parameters.4 Its context window is roughly four times larger.5 Only about seven weeks separate the two releases, which says a lot about how quickly Tencent's lab is now shipping.

The architecture details point to the cost of very long contexts as the main engineering problem. Hy4 preview uses Gated DeepSeek Sparse Attention and a technique called IndexCache, which reuses sparse indexes across layers to cut the cost of long contexts.15 The model has 78 layers. The first is a dense feed-forward layer and the other 77 are MoE layers. Each MoE layer routes tokens across 256 experts plus one shared expert, and only the top eight are active per token.9 By comparison, Hy3 used 192 experts with top-8 routing.3 The move to many experts and sparse attention fits a pattern across the industry: grow total capacity while keeping per-token compute closer to a mid-size model.

Tencent has a long record with this approach. The Hunyuan-A13B technical report describes an 80-billion-parameter MoE model with 13 billion active parameters. It has a "fast-thinking" mode for routine queries and a "slow-thinking" mode for multi-step reasoning, and the report says it matched much larger models on reasoning benchmarks.7 Hy4 preview applies the same efficiency-first idea at a much larger scale.

The coding and research claims

Tencent's messaging centers on long-running autonomous work rather than chat. The company says Hy4 preview can understand, plan, debug and verify software across extended sessions. It can also turn scattered files into documents, spreadsheets and presentations, and it can build playable game prototypes from one prompt.17

The most notable claim is about research automation. Tencent says Hy4 preview ran several Codex sessions in parallel and changed its research direction as results came in. In a post-training task on a small model, Tencent says it coordinated several evaluation targets and beat an independently operating Codex on all eight benchmarks.17 If that result holds up, it points to Tencent positioning the model as a coordinator of other coding agents, not just a code generator. That is a different and arguably more ambitious goal.

On benchmarks, Tencent reports the top open-model scores on the SWE Atlas coding trio: 64.0 on Q&A, 57.8 on test writing and 53.3 on refactoring. It also reports leading open-model results on JobBench and BioMysteryBench, and says the model slightly beats GPT-5.6 Sol on SWE-bench.9 Tencent also ran a blind internal evaluation in which 163 of its experts scored 203 engineering tasks. Hy4 preview averaged 2.99, compared with 2.92 for GLM 5.3 and 2.94 for Kimi K3.17

These numbers need careful reading. Every benchmark figure is vendor-reported and has not yet been independently reproduced.9 The internal evaluation shows only small gaps between the models, and Tencent's own staff did the grading. One early analysis noted that Tencent's materials made up most of the commentary at launch, leaving outside developers to test the claims themselves.5 Our reading: Hy4 preview has made a credible case to be in the top group of Chinese open-weight models. Claims that it leads that group should be treated as provisional for now.

There is also a reason to treat this as a real step forward in coding. An analysis of Hy3 at its July launch said it was competitive on most tasks but trailed GLM-5.2 and Western flagships in coding. It posted 78% on SWE-bench Verified and 57.9% on SWE-bench Pro.3 Hy4 preview is clearly meant to close that gap. One account, relying on Tencent's figures, calls it the strongest open-weight coding model China has produced, competitive with Kimi K3, GLM 5.3 and DeepSeek V4 Pro.9

Tencent is open about the limits

The coverage agrees on one point: Tencent says what is unfinished about the model. The company acknowledges that this early version can take longer than needed on complex questions and may over-verify its own answers.1412 Its own description says the model can over-reason and over-verify on complex tasks.17

These are not small issues for a model aimed at agent workflows. Extra reasoning steps cost money and time, and long-running tasks multiply both. The disclosure is a sensible choice, and it explains the "preview" label. Tencent's Hy3 preview in April was followed by a full release in July.2 Developers should expect Hy4 to go through a similar refinement stage before it is ready for production.

Tencent also says it trained the model on work-related data from its own software engineers, game developers, finance analysts and security experts.17 That helps explain the focus on office and engineering tasks. Tencent is building a model that reflects how its own staff work, and it can test it directly in its own products.

Talent, products and a crowded field

Reuters framed the release as part of Tencent's growing AI spending in a crowded market. It pointed to the April launch of the company's third-generation model, the first major release since Tencent hired former OpenAI researcher Yao Shunyu to lead its AI platform development.18 The reporting differs slightly on naming. Reuters calls the April model "Hunyuan 3.0."14 Other accounts describe it as Hy3 preview, the first model built on Tencent's rebuilt pre-training and reinforcement learning infrastructure.2 Tencent now uses the shorter "Hy" brand internationally and keeps the Hunyuan name in China.2

On the product side, the route from research model to revenue is fairly direct. Hy3 already powers the WeChat assistant, Yuanbao and CodeBuddy, and its API pricing was reported at about $0.18 per million input tokens and $0.59 per million output tokens.3 One analysis puts Hy4 preview at roughly a tenth of what closed frontier models charge for output.9 Tencent's bet appears to be that open weights bring in developers, and that cheap hosted access and integration with WeChat-scale products turn that interest into usage.

The open-source approach also fits a wider trend. Commentary on the launch placed it among Chinese technology companies using open releases to win over developers abroad and to compete for business customers looking for cheaper alternatives to closed systems.1916 Licensing has been part of how Tencent got there. The Hy3 preview license excluded the EU, the UK and South Korea. The full Hy3 release dropped those limits and moved to Apache 2.0, a change that mattered especially to European companies.3 Releasing a flagship preview under a clean permissive license on launch day suggests Tencent learned from that earlier friction.

The pace since launch

Tencent's Hunyuan team has kept up a steady stream of releases since Hy4 preview. In September it open-sourced AuK, a 1.5-billion-parameter model for generating and editing speech.1 On October 6 it released Prism with Fudan University and Zhejiang University. Prism is an MIT-licensed video generation framework that uses dynamic sparse attention for native 2K joint video and audio generation.10 Like Hy4's attention design, Prism targets the high cost of full attention at scale, though in a different modality.

Taken together, these releases suggest Tencent sees open publishing as central to its research strategy, not as a one-off flagship event. Hy4 preview is the headline release. The bigger story is a lab that has turned an efficiency-focused MoE line into a fast-moving, permissively licensed flagship in a matter of months. Independent benchmarks, and whether the final release fixes the over-reasoning problem, will decide whether it competes with the best models or simply catches up to them.